Gym (openAI) environment actions space depends from actual state - reinforcement-learning

I'm using gym toolkit to create my own env and keras-rl to use my env within an agent.
The problem is that my actions space changes, it depends from actual state.
For example, i have 46 possible actions, but given a certain state only 7 are available, and i'm not able to find a way to modeling that.
I've read that question open-ai-enviroment-with-changing-action-space-after-each-step
but this did not resolve my problem.
In Gym Documentation there are not instructions to do this, only an issue on their Github repo (still open).
I can't understand how the agent (keras-rl, dqn agent) pick up an action, is it randomically choosen? but from where?
Can somebody help me? Ideas?

I've handled this by just ignoring any invalid actions and letting the exploration mechanics keep it from getting stuck. Quick and simple, but likely better ways to do it.
I think the better option is to somehow set the probability of selecting that action to zero, but I've had trouble figuring out how to do that.

Related

Deep Reinforcement Learning, how to make an agent that control many machines

Good morning, Im facing a "RL" problem, which have many constraints, the main idea is that my agent will control many different machines with for example ordering them to go out for doing their missions (we don't give importance for the mission), or ordering them to enter to the depot and choosing for them the right place where they should sit (depending from constraints).
The problem is: the agent will take decision at periods of time that are defined, for each periode we know which of actions (go out, go in) are allowed. He will for example at 8oclock decide to order for 4 machines to go out, and at 14oclock decide to bring back 2 machines(with choosing for them the right place).
In literature i show many ideas which refers to BDQ, but is it recquired for my problem ? Im thinking about having actions like [chooseMachine1, chooseMachine2,chooseMachine3...chooseMachineN, goOut, goInPlace1, goInPlace2, goInPlace3, goInPlace4]. And in the code specifying the logic that depending of the period we are, i expose for the begening a number M<=N of the machines to choose (with giving 0 probability to those actions that aren't possible for the moment' if it is 14oclock you know that only the machines that are out are concerned with the agent decision'), if the agent choose Machine1, so he will access to only the possible actions from choosing it.
So, my question is, do you think that my ideas are right ? (am beginner), my idea is to make a DQN with giving the logic for the possible/impossible actions,
Do you think that a BDQ is more accurate with my problem ? like having N branchs for N machines which have the same possible actions (brach1(Machine1) : go out, goPlace1, goPlace2 ...)
If it is the case is there any implementation examples ?
If you have ressources to advise me, i will be glad of checking them.
Thank You
What would an agent navigating a maze do in case the chosen action would run it into a wall?
I think the usual approach in RL is to allow the move and than handle the result with the environment. In such a way the environment can simply make nothing happen or even give a negative reward when an action is "disallowed".
At training convergence the agent will hopefully learn to not chose ineffective actions.

Action masking for continuous action space in reinforcement learning

Is there a way to model action masking for continuous action spaces? I want to model economic problems with reinforcement learning. These problems often have continuous action and state spaces. In addition, the state often influences what actions are possible and, thus, the allowed actions change from step to step.
Simple example:
The agent has a wealth (continuous state) and decides about spending (continuous action). The next periods is then wealth minus spending. But he is restricted by the budget constraint. He is not allowed to spend more than his wealth. What is the best way to model this?
What I tried:
For discrete actions it is possible to use action masking. So in each time step, I provided the agent with information which action is allowed and which not. I also tried to do it with contiuous action space by providing lower and upper bound on allowed actions and clip the actions smapled from actor network (e.g. DDPG).
I am wondering if this is a valid thing to do (it works in a simple toy model) because I did not find any RL library that implements this. Or is there a smarter way/best practice to include the information about allowed actions to the agent?
I think you are on the right track. I've looked into masked actions and found two possible approaches: give a negative reward when trying to take an invalid action (without letting the environment evolve), or dive deeper into the neural network code and let the neural network output only valid actions.
I've always considered this last approach as the most efficient, and your approach of introducing boundaries seems very similar to it. So as long as this is the type of mask (boundaries) you are looking for, I think you are good to go.

Passing my own project on someone else - what to do? [closed]

Closed. This question is opinion-based. It is not currently accepting answers.
Want to improve this question? Update the question so it can be answered with facts and citations by editing this post.
Closed 5 years ago.
Improve this question
Often there are situations where a project is passed on someone else. And often this process is unpleasant for both sides - the new owner complains about horrible documentation, bugs and bad design. The original owner is then bothered for months with questions about the project, requests to fix old bugs etc.
I might soon be in a situation where one of my projects will be given to someone else so I can focus on my other projects. I wonder what should I do to make this transfer as smooth as possible. What i already have is a decent documentation, the code is quite good commented and i'm still improving it. Its a medium sized project, not very large but still its not something you can code in a week.
I'm looking for a list of things that should be done in order to help the future owner taking over the project and at the same time will spare me all those annoying questions like "and what does this function do, what purpose does this class have...". I know documentation is a must - what else?
Note: although my project is in C++ i believe this is a language-agnostic question. If there are things you think are specific to some language, please mention them too.
Documentation is one thing, getting it into the head of your new project owner another. IMHO this is a typical situation where "less is more" - the less documentation your colleague has to read to understand something, the better. And, of course, learning takes time - for both of you, accept it.
So
instead of writing lots of documentation, make your code self-commentatory
have all documents / source code etc. in a clean and well named folder structure
make sure your build-process is almost completely automatic
don't forget to document your deployment process, if it is not automatic, too
clean-up, clean-up clean-up!
When taking over a project, documentation is of course desirable, but even more so is a good test suite. Trying to modify a program that you have no means of testing for correctness is a nightmare.
Documentation, but on all levels:
API docs
High level architecture: What components are there, what are their relationships and dependencies
For each component, a high level description pointing to important code sections
Tutorials: If you want to do X, here's how
Data: What data does it use and how, database schemas
Idioms: If you've created some idioms within your code, explain them
And, to start, give the guy a personal introduction to all of the above in person, hopefully doing some needed change in a pair programming way
the new owner complains about horrible documentation, bugs and bad design.
I suspect that no matter what you would do, new owner will always complain about something. People are different, so something that looks easy to understand for you, will look horrible and extremely complicated for someone else.
The original owner is then bothered for months with questions about the project, requests to fix old bugs etc.
In this case you should clearly refuse to help. If you won't refuse, you'll probably end up doing someone else's job for free. If maintaining the project is no longer your job, then the new guy should fix his problem without your help. If "the new guy" can't deal with that, he isn't suitable for the job and should quit.
Its a medium sized project,
"Medium sized" compared to what? How many lines or code, how many files, how many megabytes of code?
I wonder what should I do to make this transfer as smooth as possible. What i already have is a decent documentation, the code is quite good commented and i'm still improving it.
I would handle it like this:
First, do a sweep through the entire code and:
1.1 Remove all commented out blocks of code.
1.2 Remove all unused routines and classes (I'm talking about "forgotten" routines, not parts of utility library).
1.3 Make sure all code follow consistent formatting rules. I.e. you shouldn't mix class_a, ClassA and CClassA in same app, you shouldn't use different styles for putting brackets, etc.
1.4 Make sure that all names (class, variable, function) are self-explanatory. Your code should be as self-explaining as possible - this will save you from writing too much documentation.
1.5 In situations when there is a complicated or hard to understand function, write comments. Keep them as short as possible, and post only when they are absolutely necesarry.
1.6 Try to make sure that there are no known bugs left. If there are known bugs, document them and their behavior.
1.7 Remove garbage from project directories (files that are not used in project, etc.)
1.8 If possible, make sure that code still compiles and works as expected.
Generate html documentation with doxygen. Reveiw it few times, modify code comments a bit until you're satisfied. Or until you're somewhat satisfied with the result. Do not skip this step.
If there is a version control repository (say, git repository) with entire development history, hand it over to a new maintainer, or give him(her?) a functional copy of the repository. This will be useful for (git )bisecting and finding source of the bugs.
Once it is done, and code is transferred to a new maintainer, do not offer "free help", unless you're paid for it (or unless you get something else for helping, or unless it is order from your boss which makes helping new maintainer a part of your current task). Maintaining the code is no longer your job, and if new maintainer can't handle it, he isn't qualified for the job.
I think most of the problems can be avoided with just two simple rules.
Keep the code consistent with platform style guide.
Naming, naming and naming.
If the project is huge, then you just need to run some code camps with the new guys. There's no shortcut for this one.
Remember also that complaining happens mostly because new guy is not qualified enough, i.e. doesn't understand something. That's why it is important to keep things simple. And in case he is more qualified, then I guess you deserve it ;)
Some good advice where to start hacking/changing things is always better than documentation. Consider documentation as a backup material after you are familiar with the code, it should never be the starting point (except if you are exceptional technical writer with unlimited resources and time)
If there is good documentation and commented code as you say, then you've done your part. Just make sure that the documentation includes high-level documentation (architecture, data flow, etc.) as well as lower module or procedure-level documentation.
If this is a situation where you can, I would strongly suggest you protect yourself with some type of contract that specifies what future support (if any) you will provide and for how long.
I think for a situation like this the most important thing is a working, complete build that automatically compiles, documents, and tests the project. That way, there is a well defined point at which the new developer has it working. He can then figure stuff out from the tests and documentation, in principal.

What's the best way to become familiar with a large codebase? [closed]

As it currently stands, this question is not a good fit for our Q&A format. We expect answers to be supported by facts, references, or expertise, but this question will likely solicit debate, arguments, polling, or extended discussion. If you feel that this question can be improved and possibly reopened, visit the help center for guidance.
Closed 10 years ago.
Joining an existing team with a large codebase already in place can be daunting. What's the best approach;
Broad; try to get a general overview of how everything links together, from the code
Narrow; focus on small sections of code at a time, understanding how they work fully
Pick a feature to develop and learn as you go along
Try to gain insight from class diagrams and uml, if available (and up to date)
Something else entirely?
I'm working on what is currently an approx 20k line C++ app & library (Edit: small in the grand scheme of things!). In industry I imagine you'd get an introduction by an experienced programmer. However if this is not the case, what can you do to start adding value as quickly as possible?
--
Summary of answers:
Step through code in debug mode to see how it works
Pair up with someone more familiar with the code base than you, taking turns to be the person coding and the person watching/discussing. Rotate partners amongst team members so knowledge gets spread around.
Write unit tests. Start with an assertion of how you think code will work. If it turns out as you expected, you've probably understood the code. If not, you've got a puzzle to solve and or an enquiry to make. (Thanks Donal, this is a great answer)
Go through existing unit tests for functional code, in a similar fashion to above
Read UML, Doxygen generated class diagrams and other documentation to get a broad feel of the code.
Make small edits or bug fixes, then gradually build up
Keep notes, and don't jump in and start developing; it's more valuable to spend time understanding than to generate messy or inappropriate code.
this post is a partial duplicate of the-best-way-to-familiarize-yourself-with-an-inherited-codebase
Start with some small task if possible, debug the code around your problem.
Stepping through code in debug mode is the easiest way to learn how something works.
Another option is to write tests for the features you're interested in. Setting up the test harness is a good way of establishing what dependencies the system has and where its state resides. Each test starts with an assertion about the way you think the system should work. If it turns out to work that way, you've achieved something and you've got some working sample code to reproduce it. If it doesn't work that way, you've got a puzzle to solve and a line of enquiry to follow.
One thing that I usually suggest to people that has not yet been mentioned is that it is important to become a competent user of the existing code base before you can be a developer. When new developers come into our large software project, I suggest that they spend time becoming expert users before diving in trying to work on the code.
Maybe that's obvious, but I have seen a lot of people try to jump into the code too quickly because they are eager to start making progress.
This is quite dependent on what sort of learner and what sort of programmer you are, but:
Broad first - you need an idea of scope and size. This might include skimming docs/uml if they're good. If it's a long term project and you're going to need a full understanding of everything, I might actually read the docs properly. Again, if they're good.
Narrow - pick something manageable and try to understand it. Get a "taste" for the code.
Pick a feature - possibly a different one to the one you just looked at if you're feeling confident, and start making some small changes.
Iterate - assess how well things have gone and see if you could benefit from repeating an early step in more depth.
Pairing with strict rotation.
If possible, while going through the documentation/codebase, try to employ pairing with strict rotation. Meaning, two of you sit together for a fixed period of time (say, a 2 hour session), then you switch pairs, one person will continue working on that task while the other moves to another task with another partner.
In pairs you'll both pick up a piece of knowledge, which can then be fed to other members of the team when the rotation occurs. What's good about this also, is that when a new pair is brought together, the one who worked on the task (in this case, investigating the code) can then summarise and explain the concepts in a more easily understood way. As time progresses everyone should be at a similar level of understanding, and hopefully avoid the "Oh, only John knows that bit of the code" syndrome.
From what I can tell about your scenario, you have a good number for this (3 pairs), however, if you're distributed, or not working to the same timescale, it's unlikely to be possible.
I would suggest running Doxygen on it to get an up-to-date class diagram, then going broad-in for a while. This gives you a quickie big picture that you can use as you get up close and dirty with the code.
I agree that it depends entirely on what type of learner you are. Having said that, I've been at two companies which had very large code-bases to begin with. Typically, I work like this:
If possible, before looking at any of the functional code, I go through unit tests that are already written. These can generally help out quite a lot. If they aren't available, then I do the following.
First, I largely ignore implementation and look only at header files, or just the class interfaces. I try to get an idea of what the purpose of each class is. Second, I go one level deep into the implementation starting with what seems to be the area of most importance. This is hard to gauge, so occasionally I just start at the top and work my way down in the file list. I call this breadth-first learning. After this initial step, I generally go depth-wise through the rest of the code. The initial breadth-first look helps to solidify/fix any ideas I got from the interface level, and then the depth-wise look shows me the patterns that have been used to implement the system, as well as the different design ideas. By depth-first, I mean you basically step through the program using the debugger, stepping into each function to see how it works, and so on. This obviously isn't possible with really large systems, but 20k LOC is not that many. :)
Work with another programmer who is more familiar with the system to develop a new feature or to fix a bug. This is the method that I've seen work out the best.
I think you need to tie this to a particular task. When you have time on your hands, go for whichever approach you are in the mood for.
When you have something that needs to get done, give yourself a narrow focus and get it done.
Get the team to put you on bug fixing for two weeks (if you have two weeks). They'll be happy to get someone to take responsibility for that, and by the end of the period you will have spent so much time problem-solving with the library that you'll probably know it pretty well.
If it has unit tests (I'm betting it doesn't). Start small and make sure the unit tests don't fail. If you stare at the entire codebase at once your eyes will glaze over and you will feel overwhelmed.
If there are no unit tests, you need to focus on the feature that you want. Run the app and look at the results of things that your feature should affect. Then start looking through the code trying to figure out how the app creates the things you want to change. Finally change it and check that the results come out the way you want.
You mentioned it is an app and a library. First change the app and stick to using the library as a user. Then after you learn the library it will be easier to change.
From a top down approach, the app probably has a main loop or a main gui that controls all the action. It is worth understanding the main control flow of the application. It is worth reading the code to give yourself a broad overview of the main flow of the app. If it is a GUI app, creating a paper that shows which screens there are and how to get from one screen to another. If it is a command line app, how the processing is done.
Even in companies it is not unusual to have this approach. Often no one fully understands how an application works. And people don't have time to show you around. They prefer specific questions about specific things so you have to dig in and experiment on your own. Then once you get your specific question you can try to isolate the source of knowledge for that piece of the application and ask it.
Start by understanding the 'problem domain' (is it a payroll system? inventory? real time control or whatever). If you don't understand the jargon the users use, you'll never understand the code.
Then look at the object model; there might already be a diagram or you might have to reverse engineer one (either manually or using a tool as suggested by Doug). At this stage you could also investigate the database (if any), if should follow the object model but it may not, and that's important to know.
Have a look at the change history or bug database, if there's an area that comes up a lot, look into that bit first. This doesn't mean that it's badly written, but that it's the bit everyone uses.
Lastly, keep some notes (I prefer a wiki).
The existing guys can use it to sanity check your assumptions and help you out.
You will need to refer back to it later.
The next new guy on the team will really thank you.
I had a similar situation. I'd say you go like this:
If its a database driven application, start from the database and try to make sense of each table, its fields and then its relation to the other tables.
Once fine with the underlying store, move up to the ORM layer. Those table must have some kind of representation in code.
Once done with that then move on to how and where from these objects are coming from. Interface? what interface? Any validations? What preprocessing takes place on them before they go to the datastore?
This would familiarize you better with the system. Remember that trying to write or understand unit tests is only possible when you know very well what is being tested and why it needs to be tested in only that way.
And in case of a large application that is not driven towards databases, I'd recommend an other approach:
What the main goal of the system?
What are the major components of the system then to solve this problem?
What interactions each of the component has among them? Make a graph that depicts component dependencies. Ask someone already working on it. These componentns must be exchanging something among each other so try to figure out those as well (like IO might be returning File object back to GUI and like)
Once comfortable to this, dive into component that is least dependent among others. Now study how that component is further divided into classes and how they interact wtih each other. This way you've got a hang of a single component in total
Move to the next least dependent component
To the very end, move to the core component that typically would have dependencies on many of the other components which you've already tackled
While looking at the core component, you might be referring back to the components you examined earlier, so dont worry keep working hard!
For the first strategy:
Take the example of this stackoverflow site for instance. Examine the datastore, what is being stored, how being stored, what representations those items have in the code, how an where those are presented on the UI. Where from do they come and what processing takes place on them once they're going back to the datastore.
For the second one
Take the example of a word processor for example. What components are there? IO, UI, Page and like. How these are interacting with each other? Move along as you learn further.
Be relaxed. Written code is someone's mindset, froze logic and thinking style and it would take time to read that mind.
First, if you have team members available who have experience with the code you should arrange for them to do an overview of the code with you. Each team member should provide you with information on their area of expertise. It is usually valuable to get multiple people explaining things, because some will be better at explaining than others and some will have a better understanding than others.
Then, you need to start reading the code for a while without any pressure (a couple of days or a week if your boss will provide that). It often helps to compile/build the project yourself and be able to run the project in debug mode so you can step through the code. Then, start getting your feet wet, fixing small bugs and making small enhancements. You will hopefully soon be ready for a medium-sized project, and later, a big project. Continue to lean on your team-mates as you go - often you can find one in particular who is willing to mentor you.
Don't be too hard on yourself if you struggle - that's normal. It can take a long time, maybe years, to understand a large code base. Actually, it's often the case that even after years there are still some parts of the code that are still a bit scary and opaque. When you get downtime between projects you can dig in to those areas and you'll often find that after a few tries you can figure even those parts out.
Good luck!
You may want to consider looking at source code reverse engineering tools. There are two tools that I know of:
SWAG Kit (Linux only) link
Bauhaus academic commercial
Both tools offer similar feature sets that include static analysis that produces graphs of the relations between modules in the software.
This mostly consists of call graphs and type/class decencies. Viewing this information should give you a good picture of how the parts of the code relate to one another. Using this information, you can dig into the actual source for the parts that you are most interested in and that you need to understand/modify first.
I find that just jumping in to code can be a a bit overwhelming. Try to read as much documentation on the design as possible. This will hopefully explain the purpose and structure of each component. Its best if an existing developer can take you through it but that isn't always possible.
Once you are comfortable with the high level structure of the code, try to fix a bug or two. this will help you get to grips with the actual code.
I like all the answers that say you should use a tool like Doxygen to get a class diagram, and first try to understand the big picture. I totally agree with this.
That said, this largely depends on how well factored the code is to begin with. If its a gigantic mess, it's going to be hard to learn. If its clean, and organized properly, it shouldn't be that bad.
See this answer on how to use test coverage tools to locate the code for a feature of interest, without knowing anything about where that feature is, or how it is spread across many modules.
(shameless marketing ahead)
You should check out nWire. It is an Eclipse plugin for navigating and visualizing large codebases. Many of our customers use it to break-in new developers by printing out visualizations of the major flows.

The best way to familiarize yourself with an inherited codebase

Stacker Nobody asked about the most shocking thing new programmers find as they enter the field.
Very high on the list, is the impact of inheriting a codebase with which one must rapidly become acquainted. It can be quite a shock to suddenly find yourself charged with maintaining N lines of code that has been clobbered together for who knows how long, and to have a short time in which to start contributing to it.
How do you efficiently absorb all this new data? What eases this transition? Is the only real solution to have already contributed to enough open-source projects that the shock wears off?
This also applies to veteran programmers. What techniques do you use to ease the transition into a new codebase?
I added the Community-Building tag to this because I'd also like to hear some war-stories about these transitions. Feel free to share how you handled a particularly stressful learning curve.
Pencil & Notebook ( don't get distracted trying to create a unrequested solution)
Make notes as you go and take an hour every monday to read thru and arrange the notes from previous weeks
with large codebases first impressions can be deceiving and issues tend to rearrange themselves rapidly while you are familiarizing yourself.
Remember the issues from your last work environment aren't necessarily valid or germane in your new environment. Beware of preconceived notions.
The notes/observations you make will help you learn quickly what questions to ask and of whom.
Hopefully you've been gathering the names of all the official (and unofficial) stakeholders.
One of the best ways to familiarize yourself with inherited code is to get your hands dirty. Start with fixing a few simple bugs and work your way into more complex ones. That will warm you up to the code better than trying to systematically review the code.
If there's a requirements or functional specification document (which is hopefully up-to-date), you must read it.
If there's a high-level or detailed design document (which is hopefully up-to-date), you probably should read it.
Another good way is to arrange a "transfer of information" session with the people who are familiar with the code, where they provide a presentation of the high level design and also do a walk-through of important/tricky parts of the code.
Write unit tests. You'll find the warts quicker, and you'll be more confident when the time comes to change the code.
Try to understand the business logic behind the code. Once you know why the code was written in the first place and what it is supposed to do, you can start reading through it, or as someone said, prolly fixing a few bugs here and there
My steps would be:
1.) Setup a source insight( or any good source code browser you use) workspace/project with all the source, header files, in the code base. Browsly at a higher level from the top most function(main) to lowermost function. During this code browsing, keep making notes on a paper/or a word document tracing the flow of the function calls. Do not get into function implementation nitti-gritties in this step, keep that for a later iterations. In this step keep track of what arguments are passed on to functions, return values, how the arguments that are passed to functions are initialized how the value of those arguments set modified, how the return values are used ?
2.) After one iteration of step 1.) after which you have some level of code and data structures used in the code base, setup a MSVC (or any other relevant compiler project according to the programming language of the code base), compile the code, execute with a valid test case, and single step through the code again from main till the last level of function. In between the function calls keep moting the values of variables passed, returned, various code paths taken, various code paths avoided, etc.
3.) Keep repeating 1.) and 2.) in iteratively till you are comfortable up to a point that you can change some code/add some code/find a bug in exisitng code/fix the bug!
-AD
I don't know about this being "the best way", but something I did at a recent job was to write a code spider/parser (in Ruby) that went through and built a call tree (and a reverse call tree) which I could later query. This was slightly non-trivial because we had PHP which called Perl which called SQL functions/procedures. Any other code-crawling tools would help in a similar fashion (i.e. javadoc, rdoc, perldoc, Doxygen etc.).
Reading any unit tests or specs can be quite enlightening.
Documenting things helps (either for yourself, or for other teammates, current and future). Read any existing documentation.
Of course, don't underestimate the power of simply asking a fellow teammate (or your boss!) questions. Early on, I asked as often as necessary "do we have a function/script/foo that does X?"
Go over the core libraries and read the function declarations. If it's C/C++, this means only the headers. Document whatever you don't understand.
The last time I did this, one of the comments I inserted was "This class is never used".
Do try to understand the code by fixing bugs in it. Do correct or maintain documentation. Don't modify comments in the code itself, that risks introducing new bugs.
In our line of work, generally speaking we do no changes to production code without good reason. This includes cosmetic changes; even these can introduce bugs.
No matter how disgusting a section of code seems, don't be tempted to rewrite it unless you have a bugfix or other change to do. If you spot a bug (or possible bug) when reading the code trying to learn it, record the bug for later triage, but don't attempt to fix it.
Another Procedure...
After reading Andy Hunt's "Pragmatic Thinking and Learning - Refactor Your Wetware" (which doesn't address this directly), I picked up a few tips that may be worth mentioning:
Observe Behavior:
If there's a UI, all the better. Use the app and get a mental map of relationships (e.g. links, modals, etc). Look at HTTP request if it helps, but don't put too much emphasis on it -- you just want a light, friendly acquaintance with app.
Acknowledge the Folder Structure:
Once again, this is light. Just see what belongs where, and hope that the structure is semantic enough -- you can always get some top-level information from here.
Analyze Call-Stacks, Top-Down:
Go through and list on paper or some other medium, but try not to type it -- this gets different parts of your brain engaged (build it out of Legos if you have to) -- function-calls, Objects, and variables that are closest to top-level first. Look at constants and modules, make sure you don't dive into fine-grained features if you can help it.
MindMap It!:
Maybe the most important step. Create a very rough draft mapping of your current understanding of the code. Make sure you run through the mindmap quickly. This allows an even spread of different parts of your brain to (mostly R-Mode) to have a say in the map.
Create clouds, boxes, etc. Wherever you initially think they should go on the paper. Feel free to denote boxes with syntactic symbols (e.g. 'F'-Function, 'f'-closure, 'C'-Constant, 'V'-Global Var, 'v'-low-level var, etc). Use arrows: Incoming array for arguments, Outgoing for returns, or what comes more naturally to you.
Start drawing connections to denote relationships. Its ok if it looks messy - this is a first draft.
Make a quick rough revision. Its its too hard to read, do another quick organization of it, but don't do more than one revision.
Open the Debugger:
Validate or invalidate any notions you had after the mapping. Track variables, arguments, returns, etc.
Track HTTP requests etc to get an idea of where the data is coming from. Look at the headers themselves but don't dive into the details of the request body.
MindMap Again!:
Now you should have a decent idea of most of the top-level functionality.
Create a new MindMap that has anything you missed in the first one. You can take more time with this one and even add some relatively small details -- but don't be afraid of what previous notions they may conflict with.
Compare this map with your last one and eliminate any question you had before, jot down new questions, and jot down conflicting perspectives.
Revise this map if its too hazy. Revise as much as you want, but keep revisions to a minimum.
Pretend Its Not Code:
If you can put it into mechanical terms, do so. The most important part of this is to come up with a metaphor for the app's behavior and/or smaller parts of the code. Think of ridiculous things, seriously. If it was an animal, a monster, a star, a robot. What kind would it be. If it was in Star Trek, what would they use it for. Think of many things to weigh it against.
Synthesis over Analysis:
Now you want to see not 'what' but 'how'. Any low-level parts that through you for a loop could be taken out and put into a sterile environment (you control its inputs). What sort of outputs are you getting. Is the system more complex than you originally thought? Simpler? Does it need improvements?
Contribute Something, Dude!:
Write a test, fix a bug, comment it, abstract it. You should have enough ability to start making minor contributions and FAILING IS OK :)! Note on any changes you made in commits, chat, email. If you did something dastardly, you guys can catch it before it goes to production -- if something is wrong, its a great way to get a teammate to clear things up for you. Usually listening to a teammate talk will clear a lot up that made your MindMaps clash.
In a nutshell, the most important thing to do is use a top-down fashion of getting as many different parts of your brain engaged as possible. It may even help to close your laptop and face your seat out the window if possible. Studies have shown that enforcing a deadline creates a "Pressure Hangover" for ~2.5 days after the deadline, which is why deadlines are often best to have on a Friday. So, BE RELAXED, THERE'S NO TIMECRUNCH, AND NOW PROVIDE YOURSELF WITH AN ENVIRONMENT THAT'S SAFE TO FAIL IN. Most of this can be fairly rushed through until you get down to details. Make sure that you don't bypass understanding of high-level topics.
Hope this helps you as well :)
All really good answers here. Just wanted to add few more things:
One can pair architectural understanding with flash cards and re-visiting those can solidify understanding. I find questions such as "Which part of code does X functionality ?", where X could be a useful functionality in your code base.
I also like to open a buffer in emacs and start re-writing some parts of the code base that I want to familiarize myself with and add my own comments etc.
One thing vi and emacs users can do is use tags. Tags are contained in a file ( usually called TAGS ). You generate one or more tags files by a command ( etags for emacs vtags for vi ). Then we you edit source code and you see a confusing function or variable you load the tags file and it will take you to where the function is declared ( not perfect by good enough ). I've actually written some macros that let you navigate source using Alt-cursor,
sort of like popd and pushd in many flavors of UNIX.
BubbaT
The first thing I do before going down into code is to use the application (as several different users, if necessary) to understand all the functionalities and see how they connect (how information flows inside the application).
After that I examine the framework in which the application was built, so that I can make a direct relationship between all the interfaces I have just seen with some View or UI code.
Then I look at the database and any database commands handling layer (if applicable), to understand how that information (which users manipulate) is stored and how it goes to and comes from the application
Finally, after learning where data comes from and how it is displayed I look at the business logic layer to see how data gets transformed.
I believe every application architecture can de divided like this and knowning the overall function (a who is who in your application) might be beneficial before really debugging it or adding new stuff - that is, if you have enough time to do so.
And yes, it also helps a lot to talk with someone who developed the current version of the software. However, if he/she is going to leave the company soon, keep a note on his/her wish list (what they wanted to do for the project but were unable to because of budget contraints).
create documentation for each thing you figured out from the codebase.
find out how it works by exprimentation - changing a few lines here and there and see what happens.
use geany as it speeds up the searching of commonly used variables and functions in the program and adds it to autocomplete.
find out if you can contact the orignal developers of the code base, through facebook or through googling for them.
find out the original purpose of the code and see if the code still fits that purpose or should be rewritten from scratch, in fulfillment of the intended purpose.
find out what frameworks did the code use, what editors did they use to produce the code.
the easiest way to deduce how a code works is by actually replicating how a certain part would have been done by you and rechecking the code if there is such a part.
it's reverse engineering - figuring out something by just trying to reengineer the solution.
most computer programmers have experience in coding, and there are certain patterns that you could look up if that's present in the code.
there are two types of code, object oriented and structurally oriented.
if you know how to do both, you're good to go, but if you aren't familiar with one or the other, you'd have to relearn how to program in that fashion to understand why it was coded that way.
in objected oriented code, you can easily create diagrams documenting the behaviors and methods of each object class.
if it's structurally oriented, meaning by function, create a functions list documenting what each function does and where it appears in the code..
i haven't done either of the above myself, as i'm a web developer it is relatively easy to figure out starting from index.php to the rest of the other pages how something works.
goodluck.